AYATSK
AYATSK is a classification-focused TSK model from the 2025 paper by Xue, Yang, and Wang.
Reference
G. Xue, Y. Yang and J. Wang, "Adaptive Yager T-Norm-Based Takagi–Sugeno–Kang Fuzzy Systems," in IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 55, no. 12, pp. 9802-9815, Dec. 2025, doi: 10.1109/TSMC.2025.3621346.
Mathematical Formulation
Antecedent
The paper uses Composite Exponential membership functions (CEMF) for each input feature:
The lower bound is \(1/K > 0\), which is important because AYATSK derives the adaptive Yager index from a fixed lower bound rather than from a sample-dependent minimum.
Adaptive Yager T-norm
Rule firing strengths are computed with a parameterized Yager T-norm:
In the paper, \(\lambda\) is computed once from the input dimensionality \(D\) and the membership lower bound \(\varepsilon\):
highFIS follows that policy by deriving \(\lambda\) from the number of input features and the CEMF lower bound in AYATSK models.
Defuzzification
AYATSK uses a sum-based defuzzifier to normalize rule activations:
This is implemented in highFIS by SumBasedDefuzzifier.
Consequent (first-order)
For classification:
For regression:
Output aggregation
Final outputs are weighted averages of rule consequents:
- Classification:
- Regression:
Code ↔ Paper Correspondence
| Concept | Class / Method | Notes |
|---|---|---|
| Adaptive Yager T-norm | AYATSKClassifierModel, AYATSKRegressorModel |
Default AYATSK path computes lambda from dimension and lower bound |
| Yager exponent \(\lambda\) | YagerTNorm |
Fixed after dataset-level adaptation; not learnable per rule |
| Sum-based defuzzification | SumBasedDefuzzifier |
Normalizes \(\bar{f}_r\) across rules |
| Composite Exponential MF | CompositeExponentialMF |
Provides a positive lower bound needed by the adaptive strategy |
| Estimator wrapper | AYATSKClassifier, AYATSKRegressor |
Builds input MFs and handles training/hyperparameters |
Implementation notes
AYATSKClassifierModeluses MSE loss, Adam, and zero-initialized consequents by default.AYATSKClassifierdefaults ton_mfs=3,mf_init="grid",rule_base="coco",epochs=200, andlearning_rate=0.001.AYATSKClassifierandAYATSKRegressorexposek(CEMF parameter), with required constraintk > 1.- When
batch_size=None, the AYATSK estimator uses full-batch forN < 500and0.1 * Notherwise, matching the paper’s training policy. AYATSKRegressorremains available as a framework extension, but the paper itself evaluates classification only.
Estimator wrappers
AYATSKClassifierwrapsAYATSKClassifierModeland supports classification with paper-style adaptive Yager aggregation.AYATSKRegressorwrapsAYATSKRegressorModelfor regression tasks, but this is outside the paper’s evaluated scope.- Estimators accept the usual hyperparameters:
n_mfs,mf_init,sigma_scale,random_state,epochs,learning_rate,batch_size,shuffle,validation_data, andpatience. pfrb_max_rulesexists on the shared estimator base but is unused by AYATSK.
Alignment with the paper
- The paper evaluates AYATSK on classification datasets only.
- highFIS implements the paper-strict classification path with
CompositeExponentialMF, dataset-level adaptiveYagerTNorm, MSE loss, Adam, and zero-initialized consequents. - The regressor remains supported as a framework extension, but it should not be presented as part of the paper’s experimental validation.